8

Weather Derivative
Modelling and Valuation:
A Statistical Perspective
Anders Brix, Stephen Jewson and Christine Ziehmann
Risk Management Solutions

eather derivatives are usually priced by analysing the historical outcomes

W of the underlying weather index. In this chapter we review statistical and

actuarial methods for such analyses and discuss the relevance of arbitrage
pricing. We look at reasons for trends in historical data and describe how to estimate
and remove them. Statistical methods for modelling and validating models for
weather indexes and daily temperatures are discussed, and in particular we show
that traditional ARMA time-series models are not adequate for modelling daily
temperatures. We show how dependencies between different indexes and different
locations can be modelled and we review some of the methods for risk loading of
actuarial prices.

Introduction
Weather derivatives are different from most other derivatives in that the underlying
weather cannot be traded. Furthermore, the weather derivatives market is relatively
illiquid. This means that weather derivatives cannot be cost-efficiently replicated with
other weather derivatives, ie, for most locations the bid–ask spread is too large to
make it economical to hedge a position. One of the consequences of this is that
valuation of weather derivatives is closer to insurance pricing than to derivatives
pricing (arbitrage pricing). For this reason it is important to base valuation on
reliable historical data, and to be able to model the underlying indexes accurately.
In the future, the weather derivatives market may become more liquid, and at that
point it may be possible to use other weather derivatives for hedging and thereby
derive prices from the market for at least some contracts. However, the main purpose
of this chapter is to review how weather derivatives are priced using historical data,
and to highlight some of the challenges that arise when doing so. The presentation
is statistical in its focus on the choice of models and model validation.
Before discussing the topics of detrending (removing trends from a time-series),
index and daily temperature modelling, portfolio modelling and risk loading (the
risk premuim added to the expected payoff in order to compensate the risk bearer
for taking on risk), this chapter begins with with a discussion of the relationship
between index and payoff distributions, which will be useful in the later sections.
Throughout the chapter the methods described will be illustrated with a relatively
commonly traded contract: a New York LaGuardia, May 1–September 30 cooling
degree-day (CDD) call option. Because more than 90% of the weather derivatives
currently traded are based on temperature (WRMA, 2002), the main focus of this
chapter will be on models for such indexes. Most of the index-based methods

CLIMATE RISK AND THE WEATHER MARKET

 128
W E AT H E R D E R I VAT I V E described would, however, apply to indexes based on other weather variables. The
models for daily temperatures, on the other hand, would probably not apply to any
MODELLING AND
other index types. The formulae are illustrated with some commonly used
VA L U AT I O N : A distributions in Panel 1.
S TAT I S T I C A L
Deriving the payoff distribution
PERSPECTIVE
The simplest way to understand the distribution of possible payoffs related to a
weather index is to model the distribution of the index rather than the payoff. The
reason for this is that limits on the payoff result in payoff distributions that are
mixtures of discrete and continuous distributions in the sense that they will have
discontinuities at zero and at the limit (see the following sub-section “A call option
example”). Index distributions on the other hand are usually either discrete or
continuous. This means that it is often possible to find a standard statistical
distribution which models the index in a satisfactory way. The question of finding
such a distribution is discussed in more detail in later. Alternatively, we could model
daily temperatures, derive the corresponding index distribution. This will be
discussed further in in the “Modelling daily temperatures” section.
Once we have an estimate of the index distribution we can derive the payoff
distribution of the weather derivative. Exact and approximate expressions can be
derived in many cases, but it is often simpler and faster to simulate realisations from
the estimated index distribution, and convert each simulated index into a payoff.
Nevertheless, knowing how the payoff distribution is derived theoretically from the
index distribution can be useful for model validation and understanding which
aspects of the index distribution are the most important.
In the following, I denotes the underlying index and we denote that the
distribution of I has cumulative distribution function (CDF) F. As an example we will
consider a call option on a degree-day index, but other contract types are treated in
similar ways.

A CALL OPTION EXAMPLE

We consider a call option with strike S (in degree days), limit L (in degree-days) and
tick d (in US dollars per degree-day). The payoff P of the call option is given by

P = d (min(I, L) – min(I, S)) (1)


0 I≤S
= d(I – S) S<I≤L
d(L – S) I>L

Using the tick we can convert the limit, L, into a USdollar limit L$ = d (L–S), and the
CDF G of the payoff can be expressed as:


F (S) P=0

G (P) = 
F S+
P
d  0 < P ≤ L$ (2)

1 P > L$

Because of the strike and limit, the payoff P is partly discrete with point masses at
zero and the limit L$. Between these two points the distribution is discrete or
continuous depending on whether the underlying index distribution is discrete or
continuous.
The mean of the payoff can be conveniently calculated using the limited expected
1
value (LEV) LI for the index I :

RISK BOOKS
 129
LI (m) = E min(I, m) W E AT H E R D E R I VAT I V E
m MODELLING AND
= ∫–∞
I dF (I ) + m (1 – F (m))
VA L U AT I O N : A
Here, m is the arguement of LI and it is seen that if we let m tend to infinity, LI (m) S TAT I S T I C A L
tends to the expected value of I. Taking the expectation of Equation (1) and using
PERSPECTIVE
the definition of LI (m), we see that the mean payoff is given by:

EP = d (LI (L) – LI (S))

This expression shows that the LEV function is one of the fundamentally relevant
properties of the index distribution when pricing weather derivative options. If other
moments of the payoff distribution are of interest, they can be calculated using
k
higher-order LEV functions such as E min(m, I ) .
The formulae for calculating payoff distributions and LEV functions are illustrated
with some commonly used distributions in examples 1–3 in Panel 1.

Adjusting for warming and cooling trends

One of the common requirements for accurate modelling of weather indexes is that
the series of historical indexes is stationary, which roughly means that the
distribution of indexes does not change over time. Obviously, stationarity cannot be
assumed: climate change and urbanisation may both be reflected in data. Climate
change refers to variations in the Earth’s climate on large spatial scales due to either
natural variation or anthropogenic changes in the composition of the atmosphere. In
contrast, urbanisation, which can also greatly alter a measurement station’s
temperature record, is a local and regional effect.
In global temperature trend studies, correction for local effects is often made
either by using rural stations alone or by using rural stations to “correct”
measurements from stations in urban areas. Several papers describe such studies,
and their main findings are:

1. The global average surface temperature has increased by 0.6 (+/– 0.2) degrees
Celsius since the late 19th century (IPCC, 2001).
2. Trends depend on the period chosen. For the US, for example, the temperature
history since 1910 can be divided into three periods: a warming period until 1940,
a cooling period from 1940 to 1970, and the recent warming period from 1970 to
the present (Knappenberger et al., 2001).
3. Trends depend on location. The recent period of warming has been almost global,
but the largest increases of temperature have occurred over the mid- and high
latitudes of continents in the northern hemisphere. Year-round cooling is only
evident in the northwestern North Atlantic and the central North Pacific Oceans
(although the North Atlantic cooling appears to have reversed recently, see for
example Hansen et al., 1996 and IPCC, 2001).
4. Trends for maximum and minimum temperatures are different. The diurnal range,
the difference between daily maximum and daily minimum temperatures, is
decreasing, although not everywhere. On average, minimum temperatures are
increasing at about twice the rate of maximum temperatures (see for example
Easterling et al., 1997 and IPCC, 2001).
5. A recent study of Knappenberger et al. (2001) shows that the trends are not
uniform; cool days are much warmer than they used to be, whereas warm days are
not.

Urbanisation effects are best demonstrated by comparing temperature time-series

from neighbouring stations where one is from an area with little urban change while
major changes have been made to the surroundings of the other. Such a comparison

CLIMATE RISK AND THE WEATHER MARKET

 130
W E AT H E R D E R I VAT I V E
MODELLING AND
PA N E L 1
VA L U AT I O N : A
D E R I V I N G T H E PAY O F F D I S T R I B U T I O N
S TAT I S T I C A L
PERSPECTIVE
Example 1
Consider the case where the underlying index follows a log normal distribution with
parameters µ and σ2, ie, log I follows a normal distribution with mean µ and variance
2
σ . The payoff CDF is:


 
log S – µ
Φ P=0
σ

G(P) = Φ  log (S + P / d) – µ
σ  0 < P ≤ L$

1 P > L$

where Φ is the standard normal CDF. The LEV function is:

   
2
σ log m – µ
LI (m) = exp µ + Φ – σ + m (1 – G(m))
2 σ

Hence the expected payoff is:

     
2
σ log L – µ log S – µ
E P = d exp µ + Φ –σ –Φ –σ
2 σ σ

+ dL (1 – G (L)) – dS (1 – G (S))

Example 2
2
If the index follows a normal distribution, with mean µ and variance σ , the situation
is a bit more complicated because the limited expected moments are less tractable.
The payoff CDF is simple:


 
S–µ
Φ P=0
σ

G(P) = Φ  S+P/d–µ
σ  0 < P ≤ L$

1 P > L$

The LEV function is given by

Lµ, σ(m) = σL 0, 1  m–µ

σ  + µΦ  m–µ
σ  
+m 1–Φ  m–µ
σ 
Here the LEV L0,1 for the standard normal is given by

 
m

∫
u 1 2
L0, 1(m) = exp – u du
–∞ 
2π 2

which can be calculated numerically.

RISK BOOKS
 131
W E AT H E R D E R I VAT I V E
Example 3
The general formulae above apply equally well when the underlying index follows a MODELLING AND
discrete distribution. Consider the case where I follows a negative binomial VA L U AT I O N : A
distribution. Although it is easy to calculate numerically, there is no closed-form
expression for the CDF of this distribution. In the following we will denote the CDF of
S TAT I S T I C A L
2
the negative binomial with mean r (1 – p)/p and variance r (1 – p)/p by Fr,p . The PERSPECTIVE
payoff CDF is given by Equation (2) with F replaced by Fr,p and the LEV function is
given by:

rp
Lr, p (m) = Fr+1, p (m) + m (1 – Fr, p(m))
1–p

can only be made if the two stations are in the same microclimate region. As an
example, this chapter will compare the CDD index for New York LaGuardia Airport
with the corresponding index for Central Park, New York. These two stations are in
the vicinity of each other, but whereas much has been built in the LaGuardia area
over the last 30–40 years and it is at the water’s edge, not much has changed
structurally over the same period around Central Park, in the heart of the New York
City. Figure 1 shows the historical CDD indexes for the two locations with linear
trends overlaid. Visually, the difference between the two plots is striking, and t-tests
of the significance of the slopes of the trendline reveal that the Central Park trend is
not significant (p=34%), while the LaGuardia trend is highly significant (p=0.06%).
For the purposes of this chapter, it is not important to distinguish between local
and global trends because we are interested in removing the combined trend. There
are many models in the statistical literature for estimating distributions with trends.
The way the trend is incorporated is often linear or multiplicative in the mean (for
example, an index is typically modelled as trend plus noise for a normal distribution,
and trend multiplied by noise for a log-normal distribution), and is usually chosen
on the basis of mathematical convenience rather than reflecting reality (since it can
be difficult to find a ‘realistic’ model for many applications and since the
appropriateness of the model must be validated anyway). This section separates
trend estimation from distribution estimation on the basis that this simplifies the
calculations, and because we can then treat all distributions in the same way.

1. Historical indexes for New York LaGuardia Airport (left) and Central Park, New
York (right). Linear trends have been superimposed.

CLIMATE RISK AND THE WEATHER MARKET

 132
W E AT H E R D E R I VAT I V E Trend estimation is done in essentially the same way for daily temperatures and
annually compounded indexes. Because seasonal effects can complicate the process
MODELLING AND
of daily detrending, detrending of annually compounded indexes is described first,
VA L U AT I O N : A and daily detrending later. Non-stationarity due to seasonality is discussed under the
S TAT I S T I C A L heading of “Capturing seasonality of temperatures” below.
There can be other sources of non-stationarity than trends, such as, relocation of
PERSPECTIVE
measurement stations or equipment and sudden changes to the surroundings of
stations. The effects of such measurement discontinuities can be better dealt with by
first enhancing the underlying data, as described in detail in Chapters 5 and 6. This
chapter is concerned with the more gradual changes that remain after those that are
associated with measurement discontinuities have been removed.

INDEX DETRENDING
The assumption of the trend model used here is that an index Ii can be represented
as a sum of a trend Ri and a random variable ei :

Ii = Ri + ei, i=1, ..., n

ei are assumed to be independent and identically distributed with mean zero. The
~
detrended indexes, Ii , are then defined as
~
Ii = Ii – R̂ + R̂n (3)

where R̂and R̂n are the estimated trends for indexes i and n.
In this way, the mean of all indexes are shifted to the estimated mean of the last
index. Often the contract will commence a year (or more) after of the end of the
historical indexes. If the trend is thought to continue after the last historical data
point it can be extrapolated to year n + k, where k is the number of years to
extrapolate forward. We then replace Rn by Rn+k in Equation (3), and get the k-year
ahead forward detrended indexes. For example, suppose we are looking at data from
a station which has experienced large growth in urbanisation in its surroundings in
recent years and where the urbanisation is still continuing. If we are considering a
contract for the winter, the most recent historical data would be from the previous
winter and we would need to extrapolate the estimated trend in order to capture the
trend introduced by continuing urbanisation.

Parametric trends
In this chapter, trends are assumed to be smooth and vary slowly in time since we
have assumed that jumps due to for example station relocations have been removed.
Therefore, it is often reasonable to approximate trends by parametric curves such as
linear or polynomial functions. The standard way of estimating the parameters of the
trend is by ordinary least squares (OLS), ie, minimising the sum:

∑(I – R )
i =1
i i
2

With yi denoting the year of index i, the trend Ri could then be parameterised by, for
example:

Ri = a + b y i (linear)

2
Ri = a + b yi + c yi (quadratic)

Ri = a exp(b yi ) (exponential)

RISK BOOKS
 133
OLS estimates are known to be sensitive to extreme observations, so if outliers are W E AT H E R D E R I VAT I V E
present, a more robust estimation procedure may be needed – least absolute
n MODELLING AND
deviations for example (ie, minimise ∑ | Ii – Ri |. See Huber (1981) for a detailed
i=1
treatment of robust statistics).
2 VA L U AT I O N : A
S TAT I S T I C A L
Non-parametric trends
PERSPECTIVE
Sometimes it may be desirable to use a non-parametric trend if there is reason to
believe that parametric trends do not provide a satisfactory approximation for the
period considered. In such situations various non-parametric methods are available
(see, for example, Bowman and Azzalini, 1997). The simplest is called the “moving
average” method, where the trend in year i is estimated as the average of the
neighbouring years:

∑
1
Ri = Ii+w
2w + 1 i =–w

The number of neighbouring years, 2w + 1, is usually called the “window”, and the
years may be weighted such that years closer to the base year contribute more than
years that are further away. The main disadvantage of moving average estimation is
that it does not extrapolate the trend beyond the last historical year.
An alternative that allows extrapolation is the “loess” method (Cleveland and
Devlin, 1988), which is based on local parametric regressions. Linear loess, for
example, estimates the trend for year i by weighted linear regression, with most
weight on nearby years. Loess is known to have better theoretical properties than
moving average estimation, especially close to the edges of the observation window.
Figure 2 shows estimated linear trend and loess trends for the CDD example used
in this chapter. The difference between the two trendlines ranges from –26 to 26
CDDs.

DAILY DETRENDING
Detrending of daily temperatures can be done using the methods described above.
However, local or global warming effects may have different magnitudes in different
seasons (Hansen et al., 1996), thus create different trends at different times of the
year. This is not a problem when modelling annual indexes since such seasonality
does not appear, and similarly non-parametric trends adapt to each season.
Parametric trends, on the other hand, may need to be adapted to vary by time of the
year. One way this can be done is to estimate linear trends separately for each month

2. Estimated linear (dashed) and loess (solid) trends for the CDD example. The
original index values are connected by dotted lines.

CLIMATE RISK AND THE WEATHER MARKET

 134
W E AT H E R D E R I VAT I V E of the year and interpolate to get a trend for each day. The seasonal linear trend at
day t then becomes:
MODELLING AND
VA L U AT I O N : A Rt = a + bt t
S TAT I S T I C A L
where bt is the slope on day t, which will follow an annual pattern.
PERSPECTIVE
Even if the analysis is based on annual indexes, it may be worth using daily
detrending when considering short contracts. When using index-based trends, the
estimated trend for a weekly contract may be implausibly different from the
estimated trend for the same contract the following week. This point is illustrated in
Figure 3, which shows how estimates of linear trends of average temperature vary by
week when they are estimated using only data from that week. By detrending daily
values from a period longer than the week of the contract, we use information about
the trend in the surrounding weeks, and thus get a more stable estimate.

HOW MANY YEARS OF DATA SHOULD BE USED?

One of the reasons for linear and higher-order polynomial trends is that they provide
good approximations to smooth trends over periods of a reasonable number of
years. What is a reasonable number of years? The answer varies by location, since
many trends are due to local effects, but also because of spatial variation in global
warming. Based on backtest studies it has been found that, on average over many
stations and many years, a reasonable number of years would be between 15 and 25
3
if a linear index trend is used. If more data is used the quality of the linear trend line
deteriorates due to non-linearity, and if less is used, there is too little data to get an
accurate estimate of the index trend and distribution. It must be stressed, though,
that this guideline is valid only on average, and that the number of years that is
appropriate for individual stations may lie outside this range.

VALIDATING THE TREND

So far this section has discussed how to estimate a given trend, but avoided giving
guidelines on how to chose a trend and how to validate it. One (and arguably the
best) way to do this is by graphical checks. First, a plot of the index values can give
some idea about whether a trend might be present or not. Second, after estimating
the trend, a residual plot should be made. The residuals are the difference between
the indexes and the trend, and should be scattered around zero without any
systematic variation by time, ie, the residuals must show no clear trend, no change in
spread over time and there must not be a clear tendency for the points to be mainly

3. Linear trends of average temperature at New York LaGuardia Airport for each
week of the year.

RISK BOOKS
 135
negative or positive. For all the methods outlined above, the residuals should also be W E AT H E R D E R I VAT I V E
approximately normally distributed. With only a few index values, validation can be
MODELLING AND
a difficult task, and any external information on when changes in trends could have
occurred should be used when choosing a trend estimate. For example, many VA L U AT I O N : A
temperature measurements are made at airports, or in city centres, so information S TAT I S T I C A L
about when the airport or city has grown is useful for determining when a trend
PERSPECTIVE
could have started.

Weather index modelling

Having described detrending of historical indexes in the previous section, we will
now discuss which types of statistical distributions are appropriate for modelling the
detrended indexes. Both parametric and non-parametric methods will be considered
and a discussion on model validation will follow.

NON-PARAMETRIC INDEX MODELLING

The simplest form of non-parametric distribution is the empirical (historical)
distribution of the indexes. In the actuarial literature use of this distribution is
known as “burn” analysis. However, given that we usually deal with relatively few
historical indexes, the empirical distribution can become very rugged and sometimes
have several modes (peaks). If the distribution of the index is thought to be smooth
or unimodal, then this may not be realistic. Instead the empirical distribution can be
smoothed by a process called “kernel smoothing”, whereby the probability density
function f of the index distribution taken at the point x is estimated by the expression

n ~

f̂ (x) =
1
n ∑
i =–1
1
h
kx – Ii
h 
Here, k is a probability density function (PDF), and the degree of smoothing is
determined by the bandwidth, h. The effect of kernel smoothing is illustrated in
Figure 4, where three Gaussian kernel density estimates have been superimposed on
a histogram. Whereas the choice of smoothing function k is not very critical, the
bandwidth selection is extremely important for the overall shape of the estimated
distribution: the larger h, the more smoothing is obtained.

PARAMETRIC INDEX MODELLING

Even with a large degree of smoothing the kernel distribution may put too much
weight on the historical data; for example, we may not believe in multimodality, the

4. Histogram with Gaussian kernel-smoothed densities overlaid for three different

bandwidths (40, 100 and 250).

CLIMATE RISK AND THE WEATHER MARKET

 136
W E AT H E R D E R I VAT I V E historical distribution may not look smooth enough or we may want to extrapolate
further beyond the historical observations than is possible by kernel smoothing. If so,
MODELLING AND
we can use parametric distributions such as normal, gamma or Poisson, depending
VA L U AT I O N : A on what type of index is considered. A normal distribution is often an appropriate
S TAT I S T I C A L choice for degree-day and average indexes, especially for longer contract periods.
The reason for this is the central limit theorem, which states that, under very general
PERSPECTIVE
conditions, the sum of a number of outcomes (like daily degree-days or daily
temperatures) approximately follows a normal distribution (see, for example, Casella
and Berger, 2002). Similarly, it can be shown that extreme events, such as the number
of days with temperature exceeding a high threshold, will be approximately Poisson-
distributed if they are close to independent (see for example, Coles, 2001). However,
extreme weather events tend to occur in clusters, and so are often better modelled
by a negative binomial distribution (see for example, McCullagh and Nelder, 1989).
In general, parameters of the distribution are most efficiently estimated by the
maximum likelihood method, ie, the parameter estimates are chosen to maximise the
PDF (probability mass function for discrete distributions) as a function of the
parameters. Alternatively, parameter estimates can be obtained by deriving
expressions for the moments, and matching these with empirical estimates. For the
normal distribution the two methods are equivalent and amount to calculating the
empirical mean and variance.

VALIDATING INDEX DISTRIBUTIONS

The sparsity of data with which one is often faced in weather index modelling can
make it difficult to distinguish between a bad fit due to sampling error and a bad fit
due to wrong choice of model. This makes it particularly important to estimate how
much variation can be expected due to sampling error and to apply a variety of model
checks. The following paragraphs describe a number of graphical techniques, show
how these can be made more quantitative through simulation and discuss goodness-
of-fit tests.

PP-, QQ- and LL-plots

Two traditional ways of validating the fit of a distribution are by comparing the model
PDF and CDF with the histogram and the empirical CDF, respectively. Such
comparisons are, however, not without difficulties. For histograms the number of
bins and the bin width have to be chosen, and for CDFs it can be difficult to
distinguish between the different S-shapes that these usually take. Instead, it is
customary in statistics to look at so-called PP- and QQ-plots, which contain the same
information as the CDF but allow easier comparison of distributions.
A PP-plot is a plot of the empirical CDF against the model CDF; if the model is
good, then the points should be close to a straight line with slope of one and
intercept of zero. This way, the problem of comparing S-shapes is turned into a
comparison of straight lines, which is much easier. Note that a PP-plot is also useful
for getting a qualitative evaluation of lack of fit. If the points fall on a straight line with
an intercept different from zero this indicates that the mean is wrong. If the points
fall on a straight line with slope different from one this indicates wrong variance. The
left panel of Figure 5 shows a PP-plot for the CDD example for New York.
Because all CDFs start at zero and end at one, points in the tails of the
distributions on the PP-plot will tend to be close to a straight line. This makes it
difficult to evaluate the quality of the fit to the tails of the distribution. A QQ-plot,
where model quantiles are plotted against empirical quantiles, is a way around this
problem. Again the points should lie close to the straight line with slope of one and
intercept of zero if the fit is good. The advantage of QQ-plots is that a bad fit in the
tails of the distribution will show more clearly. The middle panel of Figure 5 shows
a QQ-plot for the CDD example.
While PP- and QQ-plots provide good tools for evaluation of the overall fit, it is

RISK BOOKS
 137
W E AT H E R D E R I VAT I V E
5. Validation plots for a normal distribution for the CDD example
MODELLING AND
VA L U AT I O N : A
S TAT I S T I C A L
PERSPECTIVE

often useful to check the fit of specific characteristics of the distribution that are
more closely related to the actual purpose of the modelling. Since we are interested
in modelling payoffs, it is important to capture accurately those features of the index
distribution which affect the basic characteristics of the payoff distribution. Hence, if
we are modelling options or swaps, we should compare the model and empirical
LEV functions. As before, we prefer to look for straight lines when validating the fit,
so we plot the two against each other. The plot created this way is called an LL-plot,
and is shown in right panel of Figure 5 for the CDD example.

Envelopes
The validation plots described in the previous section are purely descriptive, and it
can sometimes be hard to see when deviations from a straight line are due simply to
sampling variation. In order to make them more quantitative we can add confidence
intervals to the plots. It is usually not possible to derive exact expressions for such
confidence intervals – we use simulation envelopes instead. This is done by
simulating samples of the same length as the historical data from the estimated
model; if the model is good, then the CDF of the samples should fall around the
historical CDF.
Simulation envelopes for the PP-plot are obtained as follows:

Step 1. Simulate a sample of the same length n as the historical data from the
estimated model.
Step 2. Calculate the empirical CDF from the sample at each historical index.
Step 3. Repeat Steps 1 and 2 K times and store the results.
Step 4. Sort the calculated CDF at each value of the historical indexes.

To achieve 90% confidence intervals we could simulate, say, K=100 samples and pick
4
out the fifth lowest and the fifth highest CDF values for each historical index. The
envelopes created this way are pointwise confidence intervals; there is 90%
probability that the historical CDF at a given point will fall within the envelopes.
Because points on the simulated CDFs are highly dependent this does not mean that
the probability of the full historical CDF falling within the envelopes is 90% to the
power of n.
Similarly for the QQ-plot and the LL-plot, we calculate the quantiles
corresponding to the quantiles of the historical indexes, and LEV at each historical
index. (Figure 10 shows an example of a QQ-plot within envelopes.)

GOODNESS-OF-FIT TESTS
Apart from graphical checks several goodness-of-fit tests are available, of which the
2
most common are chi-square (x ), Shapiro–Wilks, Kolmogorov–Smirnov and
Anderson–Darling. Because of the small number of historical indexes that are usually

CLIMATE RISK AND THE WEATHER MARKET

 138
W E AT H E R D E R I VAT I V E available, none of these tests are very powerful, ie, they will rarely reject a bad model.
Graphical checks, such as the ones described above, often give a much better idea of
MODELLING AND
how appropriate the model is.
VA L U AT I O N : A
S TAT I S T I C A L ❏ The chi-square test can be used for checking the validity of any distribution. The
test is done by grouping the observations into intervals and comparing the
PERSPECTIVE
expected number of observations in each interval with the observed number.
Because the test relies on grouping the observations, it requires a substantial
number of observations in order to give a powerful result (a general rule of thumb
for when to use this test is that the expected number of observations in each
group must be at least five).
❏ The Shapiro–Wilks test can be used only for normal distributions (and log-normal
by transforming the observations with the logarithm). The test is quite powerful
and gives a good indication of whether a normality assumption is reasonable.
❏ Kolmogorov–Smirnov is a classical test which compares the empirical CDF with
the model CDF using the maximal vertical difference between the two. The test is
not very powerful, but should it result in a low test probability then there is good
reason not to rely on the model.
❏ Anderson–Darling can be considered as a modification of the Kolmogorov–
Smirnov test which compares the CDFs over the whole range of the distribution.
In contrast to the Kolmogorov–Smirnov test, the Anderson–Darling test
probability is model-specific and hence the test is more powerful. Both this test
and the Kolmogorov–Smirnov test can be used only for continuous distributions.

Modelling daily temperatures

The most common approach to analysing weather derivative contracts is to fit a
distribution to the detrended annual indexes. However, often we have only a few
historical values from which to estimate this distribution, resulting in significant
estimation uncertainty. This uncertainty could be reduced if the data were used more
efficiently, and the index approach has some inefficiencies, as shown by the following
examples:

❏ For a US cooling degree-day (CDD) index, the index approach uses only
information about how far above 65°F the temperature is. It does thus not
distinguish between days where the temperature is far below 65°, and days where
the temperature is just below 65°.
❏ Event indexes only use data from days on which events occurred. Data on all other
days is discarded.
❏ One-week contracts only use data for that week of the year. Data from other weeks
is discarded.
❏ For some indexes, in particular indexes relating to short periods and extreme
events, it may not be possible to find a suitable model for the index distribution.

These problems could be alleviated if the underlying daily temperature distribution

could be modelled. Furthermore, a temperature modelling approach would make it
easier to include forecasts in the weather derivative pricing process (see Chapter 10).
To see how much more efficiently data can be used in a daily model compared
with an index model, consider the May–September CDD contract for New York
LaGuardia Airport. We estimate the index distribution from a daily temperature
model and an index model using the same amount of historical data. Both estimates
of the index distribution will have an error due to estimation uncertainty, and these
errors can be quantified by looking at confidence intervals (envelopes) around the
estimated CDFs. Figure 6 shows the estimated CDFs with 90%-confidence envelopes
using a normal index distribution and a daily temperature model, respectively.

RISK BOOKS
 139
The daily model used for Figure 6 is the CJB model to be described later, but it W E AT H E R D E R I VAT I V E
must be noted that almost any model for daily temperatures would show this
MODELLING AND
apparent decrease in estimation uncertainty. This is because a daily model uses the
data more efficiently than an index model and because the comparison is based on VA L U AT I O N : A
the assumption that the model is correct (the simulation envelopes are created using S TAT I S T I C A L
the estimated model for both the index and the daily model). However, a bad daily
PERSPECTIVE
model could also produce significant bias, and hence reduce the value of increased
estimation accuracy. For this reason it is extremely important to validate a daily
model before using it for weather derivatives pricing; validation methods are
discussed in detail later.
Another reason for a thorough validation of daily validation models is that a daily
model may give a different index distribution from that given by the historical index
values (and from a normal distribution). Since a daily model is using data more
efficiently than an index model we can place more weight on the daily model results,
but only if it has been thoroughly validated.
Statistical modelling of daily temperatures has been a research subject for
decades (see for example the review in Wilks and Wilby, 1999) and, more recently, a
number of papers on weather derivative pricing that propose models for temperature
have been published (Alaton et al. (2002), Brody et al. (2002), Caballero et al.
(2002), Cao and Wei (2000), Davis (2001), Diebold and Campbell (2001), Dischel
(1998), Moreno (2000) and Torró et al. (2001)). Common to all of the above
referenced papers except Brody et al. (2002) and Cabellero et al. (2002) is that they
use ARMA-type models which will be discussed in more detail in the following
subsections.
We will start by discussing the basic steps of building a model for daily
temperatures and then show how statistical validation techniques are applied. The
model validation will show that ARMA-type models fail to validate well on two points:

1. The modelled auto-correlation function decays too quickly relative to reality.

2. The residuals deviate markedly from their theoretical distribution.

For this reason we also show validation results for the model discussed in Caballero
et al. (2002) which, in general, validates much better than ARMA models for daily
temperatures. We conclude the section with a brief discussion of some of the
problems that still remain to be solved and which, to the authors’ knowledge, apply
to all published daily temperature models.

6. Potential gain in accuracy from daily modelling over index modelling. Normal
distribution index CDF (left) and index CDF derived from CJB model (right), both
with 90%-envelopes (dotted).

CLIMATE RISK AND THE WEATHER MARKET

 140
W E AT H E R D E R I VAT I V E CAPTURING SEASONALITY OF TEMPERATURES
One basic principle underlying statistical models for stochastic processes is that the
MODELLING AND
data should be stationary, ie, the distribution should be invariant over time. Apart
VA L U AT I O N : A from trends, daily temperatures have some obvious non-stationarities, namely those
S TAT I S T I C A L due to seasonal variation.
Seasonal variation undoubtedly affects daily temperature in many complicated
PERSPECTIVE
ways (eg, in all the moments), and it will probably never be possible to remove all
types of seasonality, let alone to check that it has been done. In the following it is
assumed that most seasonality is due to seasonal variation in the mean and the
standard deviation. In mathematical terms we assume that the time-series of daily
temperatures can be decomposed as follows

Ti = mi + si T ’i (4)

where Ti is the temperature, mi is the mean temperature, si is the standard deviation

and T’i is a mean zero variance one variable that is known in meteorology as the
“anomaly” for day i. The anomalies T’i are assumed to be a stationary time-series, for
which a large number of candidate models exist.
Like estimation of trends, seasonality estimation can be either parametric or non-
parametric. The following section discusses a number of such methods for removing
the seasonality in the mean. The seasonal variance can be estimated and removed by
exactly the same method applied to the square of the difference between the
temperature and the seasonal mean. For this reason we only describe the methods
for estimating the seasonal mean.

Non-parametric approaches
The simplest way to estimate the seasonal mean is to average each day of the year
over the historical data period, ie, the estimate for January 1 is the average of January
1 in all years and so on. When using this method, special care must be taken of leap
days. If the seasonal mean estimated this way is considered too ragged it can be
smoothed, by means of kernel smoothing, for example. One problem is that even
with smoothing, results are often too jumpy.

Parametric approaches
Another way of estimating the seasonal mean is by parameterising the mean
temperature using trigonometric functions. This has the advantage that leap days are
easy to take account of. We can estimate the mi using ordinary regression with mi
given by

mi = α cos
2π

365.25
i+ω  (5)

where α is the amplitude and ω is the phase of the seasonal mean.

Alternatively we can estimate mi in the frequency domain. In both cases more
harmonics can be included in the parameterisation of mi if necessary (ie, using
cosines to produce

 
K

∑α
2kπ
mi = k cos i+ω
k=1 365.25

While the parametric form in Equation (5) is very simple, it can be justified by plots
of power spectra of temperature anomalies which show surprisingly clear peaks at
365.25 days (and its harmonics).
The result of applying a seasonal linear trend and parametric seasonal cycles for
mean and standard deviation with two harmonics to the New York LaGuardia Airport

RISK BOOKS
 141
example can be seen in Figure 7, which shows how the detrended daily temperatures W E AT H E R D E R I VAT I V E
are decomposed into seasonal mean, seasonal standard deviation and anomalies.
MODELLING AND

ANOMALY MODELLING VA L U AT I O N : A
The biggest challenge in modelling daily temperatures is to find an appropriate S TAT I S T I C A L
model for the anomalies. We have seen that the index distribution’s LEV function
PERSPECTIVE
plays an important role in deriving the moments of the payoff distribution. Similarly
there are functions of daily temperatures which are important for capturing
moments of the index distribution from a model for daily indexes. For example, it is
important to model both the seasonal cycle and the autocorrelation of temperatures
accurately in order to get a good estimate of the index mean and variance of degree-
day or average-temperature indexes. Consider, say, a CDD index. Using the
representation in Equation (4), and assuming that the daily temperature never falls
below 65°F, we have the following expressions for the mean and variance of the
index:

n
EI = ∑m – n65
i=1
i

7. Detrended daily average temperature, seasonal mean, seasonal standard

deviation and anomalies.

CLIMATE RISK AND THE WEATHER MARKET

 142
n n
W E AT H E R D E R I VAT I V E
MODELLING AND
VI = ∑ ∑s s γ (i – j)
i=1 j=1
i j (6)

VA L U AT I O N : A Here, n is the number of days in the contract and γ is the autocorrelation function
S TAT I S T I C A L (ACF) of the anomalies. From Equation (6) it can be seen that we would under-
estimate the index variance if the model ACF, γ, decays too quickly to zero.
PERSPECTIVE
What kind of ACF behaviour do daily temperatures show? Dependencies in day to
day temperature arise from complicated atmospheric, oceanic and land surface
processes, many of which evolve slowly. In particular, ocean circulation can have
cycles from years to decades and even centuries. It would thus not be surprising if
temperature were to show dependencies on long timescales. As we shall see later, the
ACF of temperature anomalies does indeed show relatively slow decay.

NON-PARAMETRIC TIME-SERIES MODELLING

Non-parametric time-series modelling can be done by resampling from the observed
anomalies. The resampled time-series can then be used to estimate distribution
statistics of interest, weather derivative indexes, for example. Resampling should not
be done completely at random, since we would then lose all dependencies in the
time-series. Instead we could resample blocks of anomalies and concatenate these. If
the blocks are sufficiently long and if we have enough data, this should result in
simulated time-series which would mimic the behaviour of the original time-series.
The length of the blocks should be chosen so that the statistics of interest are not too
affected by the “breaks” between blocks. For example, block lengths could be chosen
to equal the contract length in the case of weather derivatives. However, it may be
difficult to capture the distribution of a time-series by resampling, even if long blocks
are used. See Davison and Hinkley (1997) for a survey of non-parametric resampling
methods.

PARAMETRIC TIME-SERIES MODELS

Traditional time-series models are parametric models, and the simplest example is
probably a first order autoregressive (AR(1)) process. For such a process the
temperature at one day is given by a linear dependency on the temperature the
previous day with some random noise added:

T’i = βT’i–1 + εi

Here T’i denotes the anomaly at day i and the εi are independent and identically
distributed random variables following a mean zero normal distribution. Despite its
simplicity an AR(1) process can be a useful model for a variety of problems, but is
unfortunately too simple for daily temperature anomalies (one reason being that the
ACF decays too quickly). A simple extension of the AR(1) process provides us with a
flexible class of models, known as ARMA models, which can be used to approximate
5
any stationary time-series model. The definition of an ARMA( p, q) process is:

T’i = φ1 T’i–1 + ... + φp T’i–p + θ1 εi–1 + ... + θi–q εi–q + εi

The interpretation of the model is that the temperature today depends in a linear way
on the temperatures on the previous p days through the parameters φ1, ..., φp. Just as
in the AR(1) case random pertubations are added to reflect the fact that we do not
expect the temperature today to be a perfect linear function of the past p days’
temperatures. The θ1, ..., θq are parameters used to express linear dependence
between the random pertubations.
Although ARMA models can, in theory, approximate any time-series model, it is
actually not the best class of models for temperature anomalies. This is partly because
of the slow decay of the ACF of temperature anomalies (see Figure 8 and the

RISK BOOKS
 143
subsection on validation using ACFs), which by Equation (6) implies that an ARMA W E AT H E R D E R I VAT I V E
model would result in underestimated variances for degree-day indexes – a fact that
6 MODELLING AND
has been recognised by many participants in the weather market for some time. If
we were to model a slower decay of the ACF using an ARMA model, we would need VA L U AT I O N : A
far more parameters than would be feasible to estimate in practice. A more S TAT I S T I C A L
parsimonious choice of model (in the following referred to as the CJB model) that
PERSPECTIVE
preserves much of the ARMA model flexibility but has slower decaying ACF is
described in Caballero et al. (2002).
In the following sections we illustrate classical statistical methods for model
validation, which provide more evidence on how ARMA models fail to provide
adequate modelling of daily temperatures and how the CJB model overcomes many
(but not all) of the problems associated with ARMA models.

Validation of daily temperature models

Modelling daily temperatures is a difficult task and careful validation of models must
be undertaken before putting them into practical use for pricing weather derivatives.
In the following subsections we show some simple tools for evaluation of daily
temperature models for weather derivative pricing. The methods are illustrated on
the New York example, using an ARMA(3,1) model and the CJB model.

AUTOCORRELATION FUNCTION
Equation (6) and the discussion following it highlighted the importance of capturing
the ACF of the temperature anomalies, making the comparison of the model ACF
with the empirical ACF an appropriate step in validation of the model. However,
since all ACFs start at one and usually decay quickly over the first few lags,
comparison can be difficult using the usual representation of an ACF. Instead, we
compare ACFs by plotting the logarithm of the ACF against logarithm of the lag. The
plot in Figure 8 is a plot of this type and emphasises what we have already
mentioned: that the ARMA model ACF decays too fast to capture the behaviour of the
empirical ACF.

DISTRIBUTION OF RESIDUALS
Residuals are the difference between the observed anomalies and the predicted
anomalies based on the past observations, ie, the historical one-step prediction
errors. The parameters used for the prediction are estimated from the full time-
series. For most models, including the two considered here, it is possible to derive
theoretical expressions for the distribution of the residuals. We can thus check the

8. Comparison of ACFs on a log–log scale: empirical ACF (circles), ARMA(3,1)

(dashed) and CJB model (unbroken).

CLIMATE RISK AND THE WEATHER MARKET

 144
W E AT H E R D E R I VAT I V E time-series model by comparing the residuals from the model with their theoretical
distribution. For non-parametric methods the theoretical distribution of the residuals
MODELLING AND
is unknown and hence cannot be checked.
VA L U AT I O N : A The left plot of Figure 9 shows a residual QQ-plot for the ARMA(3,1) model, and
S TAT I S T I C A L the right plot shows a QQ-plot for the CJB model. We see that whereas the CJB model
shows consistency with the historical anomaly distribution, the ARMA(3,1) model
PERSPECTIVE
results in a residual distribution with far too light tails.

DISTRIBUTION OF ANOMALIES
The next step in validation is the verification of the distribution of the anomalies. For
mathematical convenience most time-series models assume a normal distribution, so
a natural first step is to compare the empirical anomaly distribution with that of a
normal through PP- and QQ-plots. If a normal distribution is not suitable, it is often
possible to find a transformation that makes the anomaly distribution approximately
normal.
Models also exist for time-series with non-normal behaviour such as heavy-tailed
or skewed distributions. For such models, however, it can be difficult to verify when
the model is stationary and determine exactly what the marginal distributions are.

ACCURATE MODELLING OF WEATHER DERIVATIVE INDEXES

While it is important that the time-series of temperatures is modelled accurately, the
ultimate goal is to model weather derivative indexes. The transformation of
temperatures to a degree-day or an event-specific index (such as a critical-day index)
is non-linear, and hence it is not clear how small model deficiencies at the
temperature level will propagate to errors in the index distribution. This, however, is
not a problem for a period-average temperature index. We need to validate the index
distribution, derived from a daily model, using the same tools as we used for
validating the index model. As an example, the two plots in Figure 10 show the QQ-
plots for the index distribution resulting from using an ARMA(3,1) and the CJB
model.
Imposing 90%-confidence envelopes on the plots we see that the CJB model falls
within the envelopes, but that the ARMA model is questionable because of the
relatively large number of points that fall outside or on the boundary of the
envelopes.

Some outstanding issues in daily temperature modelling

We have seen that ARMA models are not adequate for daily temperature modelling
and that the CJB model is a better alternative. However, there are still some

9. QQ-plot for the residuals using an ARMA(3,1) model (left) and CJB model
(right) for New York LaGuardia Airport.

RISK BOOKS
 145
W E AT H E R D E R I VAT I V E
10. QQ-plot for the index distribution for New York LaGuardia Airport using an
ARMA(3,1) model (left) and CJB model (right). MODELLING AND
VA L U AT I O N : A
S TAT I S T I C A L
PERSPECTIVE

unresolved issues which means that even the CJB model may not be adequate in all
situations. In particular some locations show strong seasonal variation beyond that
of mean and standard deviation, which in turn means that both ARMA and CJB
models validate badly for these locations.
Another issue is that modelling daily temperatures for several locations
simultaneously is complicated by the fact that the spatial correlation structure is non-
stationary. This makes it difficult to exploit all the information in daily temperatures
and as a result dependencies between locations are usually modelled on an index
level instead (see the following section).

Capturing index dependencies

Weather indexes often show strong geographical dependencies – a fact that must be
captured by our modelling approach. One reason for this is that pricing of a weather
contract is generally done in the context of the current portfolio: the more risky the
portfolio gets by adding the contract, the less that contract is worth. Furthermore,
portfolio modelling is essential to valuing a book of contracts and the principles for
risk loading described in the following section rely crucially on appropriate joint
modelling of the contracts in the portfolio. To model dependencies, various tools
exist – mostly based on correlations. However, correlations are not just correlations!
The usual notion of correlation is linear correlation, which is what is calculated
in spreadsheets and other standard software packages. Linear correlation is only an
appropriate measure of dependency if the joint distribution of the indexes is close
to Gaussian or elliptical (Embrechts et al., 2000). If this is not the case, linear
correlations may fail to accurately capture the dependency between indexes,
especially in the tails of the distribution. Moreover, the correlation coefficient need
not have a maximum of one and a minimum of minus one in the non-Gaussian case.
Instead the minimum and maximum linear correlation will depend on the
distribution and its parameters, which makes it more difficult to evaluate the degree
of dependence. Fortunately other methods for capturing dependence exist, two of
which are discussed here: rank correlation and copulas.

RANK CORRELATION
Rank correlation is a alternative way of measuring dependencies between
distributions. All values between one and minus one are possible whatever the
distribution, which makes interpretation of rank correlation easier than linear
correlation in the non-Gaussian case.
Rank correlation is calculated by transforming all indexes by their empircal CDF
and then computing the linear correlation between the transformed indexes. This is
equivalent to computing the linear correlation between the rank of the sorted

CLIMATE RISK AND THE WEATHER MARKET

 146
W E AT H E R D E R I VAT I V E indexes (hence the name) and the method can be seen as a non-parametric approach
to capturing dependence between indexes.
MODELLING AND
VA L U AT I O N : A COPULAS
S TAT I S T I C A L The notion of copulas is based on the same fundamental idea as the rank
correlations, namely that dependency between distributions is compared after the
PERSPECTIVE
distributions have been transformed to uniforms: a copula is simply the CDF of a
multivariate distribution with uniform marginals. For example the Gaussian copula is
the CDF given by transforming each variable of a multivariate Gaussian distribution
by its CDF. Rank correlations can be thought of as a special type of copula, where the
dependence is given by the empirical correlation of the observed indexes
transformed to uniforms. In general, however, a copula is given as a parametric CDF,
and the parameters of the copula must either be estimated from data or guessed from
intuition. Guessing is usually not a good idea, but due to lack of data it may be
necessary. Likewise the preference of one specific copula over another can be hard
to justify from data, and is often made on the basis of mathematical convenience.

Risk loading principles

Every risk has its price, and the difference between this and a fair price is called “risk
loading”. How is an appropriate risk loading calculated? The principles outlined
below are general principles for how to obtain a price of risk based on risk measures
such as standard deviation and quantiles. Most of the principles can also be adapted
to other risk measures.
The price of a contract does not always depend directly on how risky it is. If, for
example, we know that we can buy a certain contract from somebody else in the
market at a given price it would be risk-free to offer the contract at a higher price.
This way of determining a price is driven by the market’s perception of risk rather
than one’s own. This is discussed further in the paragraphs that follow.

SIMPLE ADDITIVE RISK LOADING

The simplest risk loaded price, Pr , of a contract is the expected outcome plus λ times
the risk, where λ is a risk factor defining the speculator’s appetite for risk. Ignoring
discounting, the pricing formula is:

Pr = EP + λR (7)

where R is the risk measure and EP is the expected payoff. A measure of risk is
needed, and it is customary to use standard deviation or variance of the payoff for
this purpose. Alternatively we could use the difference between two quantiles such
as the median and the 5% quantile, whereby we get a risk measure which is more
7
akin to Value-at-Risk.
Equation (7) also allows the pricing of the contract against the full portfolio. To
do this we change the risk measure R to be a measure for the additional risk that is
taken on by trading the contract. Changes in any of the risk measures above could be
used for this purpose. For example, we could use the difference between the
standard deviation of the portfolio payoff with and without the contract as risk
measure R.

INVESTMENT EQUIVALENT PRICING

Most companies will have ways of allocating capital to contracts based on their own
internal measures of risk. Had the capital not been allocated to a contract it could
have been invested in other assets. This gives the speculator another way of pricing
a contract: the return on the allocated capital should be at least what would be
expected on an investment in a similar risk elsewhere. This is the idea behind so-

RISK BOOKS
 147
called Risk Adjusted Return On Capital (RAROC) (Nakada et al., 1999) and W E AT H E R D E R I VAT I V E
investment equivalent reinsurance pricing (Kreps, 1999).
MODELLING AND
Consider the case where an option is sold at a price Pr. The price consists of the
discounted expected payoff and risk loading, L: VA L U AT I O N : A
S TAT I S T I C A L
Pr = EP/(1 + rf ) + L (8)
PERSPECTIVE
where rf is the risk-free interest rate. In addition to the premium income we allocate
an amount of capital A to the contract and invest Pr + A in risk-free securities. The
idea is now that the average return on allocated capital should equal the return from
an investment in an alternative instrument with the same risk. The expected Return
on Allocated Capital (RAC) is thus defined by

(1 + RAC) A = (1 + rf ) (Pr + A) – EP

Combining this with Equation (8) we get the following expression for the loading

L = (RAC – rf )/(1 + rf )A

Using the estimated payoff distribution we can get an idea of how risky the option is
and use this to find a reasonable target RAC which, in turn, gives the risk loading.
Several variations of this theme are possible, see Kreps (1999).

Other topics
We have seen in this chapter that valuation of weather derivatives is a broad subject
where many factors must be taken into account. While we have tried to cover the
majority of topics there are still some that we have left out and many which could
have been further elaborated.

ARBITRAGE PRICING
Arbitrage pricing is the standard way of pricing financial derivatives in a liquid
market. As noted in this chapter’s Introduction most weather derivative contracts are
not yet liquid enough to justify arbitrage pricing but some are now traded several
times a day. This makes it worthwhile to consider how arbitrage pricing for weather
derivatives may be done.
The basic principle of arbitrage pricing is that the cost of a derivative is the cost
of creating and managing a portfolio which replicates the payoff of the derivative
contract at maturity. The active management of the replicating portfolio is what is
usually called dynamic hedging. However, the underlying index of a weather
derivative cannot be traded and hence weather derivatives cannot be hedged this
way. Instead we could replicate the payoff of a weather derivative using other
weather derivatives. While this is possible in principle, the illiquidity of the current
weather market makes it prohibitively expensive for most types of contracts.
However, should the market become liquid enough, several arbitrage strategies
would be possible.
One strategy would be to use properties of forecasts to derive a theory for how
the market would price a weather swap. Since swaps can be used to hedge options,
this gives us a way of obtaining an arbitrage price for an option. The ideas behind
this are discussed in detail in Stephen Jewson’s contribution to the End Piece.
Some authors advocate hedging of weather derivatives with other derivatives such
as power or gas derivatives (Geman, 1999). However, such hedges are not likely to
be complete hedges and we are thus left with a basis risk. In order to find the price
of the basis risk we must model the weather derivative and the underlying of the
alternative hedge jointly, for example using methods similar to those previously
discussed in this chapter. Note also that whereas HDDs are highly correlated with gas

CLIMATE RISK AND THE WEATHER MARKET

 148
W E AT H E R D E R I VAT I V E consumption, they are usually very little correlated with gas prices, which is one of
the reasons for trading weather derivatives in the first place.
MODELLING AND
VA L U AT I O N : A OTHER WEATHER VARIABLES
S TAT I S T I C A L Another challenging topic is modelling of daily weather variables other than
temperature, such as of wind speed or precipitation. These relatively uncommon
PERSPECTIVE
underlyings (compared with temperature) are dealt with thoroughly in Chapter 3.

EXTREME VALUE THEORY

One subject that has not been mentioned is the use of extreme value theory (EVT)
for evaluating risk and estimation of the distribution of extreme indexes. While this
is an interesting point of discussion, the current state of the market is such that
weather derivatives are mostly written on non-extreme risk, and as such EVT is less
applicable.

Conclusion
This chapter has described how weather derivatives can be valued by statistical and
actuarial methods based on historical data. It has given an overview of all the
important topics: estimation and adjustment of trends, modelling and validation of
detrended historical weather indexes and daily temperatures, accounting for index
dependencies and risk loading of expected payoffs.
As trading liquidity increases for certain contracts, aspects of arbitrage pricing will
become more important and in some cases replace actuarial methods. However, for
most contracts, the actuarial methods described in this chapter will continue to be
the main, and only reasonable, valuation approach that can be used.

1 The LEV function is often used in insurance for calculating mean loss to an excess
of loss reinsurance layer.
2 By “robust” we mean methods which, at the cost of accuracy, are less affected by
outliers than traditional methods.
3 Based on the authors’ own study using 50 years of data from 200 US stations.
4 In statistical literature the observed sample is often included in the K
simulations since the hypothesis modelled is a sample from the correct model. In
practice this distinction does not make a lot of difference.
5 It can be shown that for any autocovariance function γ (·) such that h→∞ϒ(h) =0,
and any integer k>0 it is possible to find an ARMA process with autocovariance
function (·) such that (h) = γ (h), h=0,1,...,k.
6 This is the authors’ own experience based on discussions with the major market
participants over the last three years.
7 Value-at-Risk (VAR) is a quantile (typically 5%) in the modelled profit and loss
distribution over a specified time period, and is one of the most common measures
of risk in finance.

BIBLIOGRAPHY

Alaton, P., B. Djehiche and D. Stillberger, 2002, “On Modelling and Pricing Weather Derivatives”,
Applied Mathematical Finance, Forthcoming.

Bowman, A, and A. Azzalini, 1997, Applied Smoothing Techniques for Data Analysis, (Oxford
University Press).

Brockwell, P., and R. Davis, 1991, Time Series: Theory and Methods, (New York: Springer-Verlag).

RISK BOOKS
 149
Brody, D. C., J. Syroka and M. Zervos, 2002, “Dynamical Pricing of Weather Derivatives”, Quantitative W E AT H E R D E R I VAT I V E
Finance, Forthcoming. MODELLING AND
Caballero, R., S. Jewson and A. Brix, 2002, “Long Memory in Surface Air Temperature: Detection, VA L U AT I O N : A
Modelling and Applications to Weather Derivative Valation”, Climate Research, Forthcoming.
S TAT I S T I C A L
Cao, M., and J. Wei,. 2000, “Equilibrium Pricing of Weather Derivatives”, Research Paper, Deptartment PERSPECTIVE
of Economics, Queen’s University, available at: http://qed.econ.queensu.ca/pub/faculty/cao/papers.html.

Casella, G., and R. L. Berger, (2002), Statistical Inference, (Pacific Grove, CA: Brooks/Coles
Publishing).

Cleveland, W., and S. Devlin, 1988, “Locally-Weighted Regression: An Approach to Regression Analysis
by Local Fitting”, Journal of the American Statistical Association, 83, pp. 596–610.

Coles, S., 2001, An Introduction to Statistical Modelling of Extreme Values, (New York: Springer
Verlag).

Davis, M., 2001, “Pricing Weather Derivatives by Marginal Value”, Quantitative Finance, 1, pp. 1–4.

Davison, A., and D. Hinkley, 1997, Bootstrap Methods and Their Application, (Cambridge University
Press).

Diebold, F. X., and S. D. Campbell, 2001, “Weather Forecasting for Weather Derivatives”, Working
Paper, Department of Economics, University of Pennsylvania, available at:
http://www.econ.upenn.edu/Centers/pier/Archive/01-031.pdf.

Dischel, R., 1998, “Black-Scholes Won’t Do”, Risk Magazine and Energy and Power Risk Management
Report on Weather Risk, http://www.financewise.com/public/edit/energy/weatherrisk/wthr-options.htm.

Easterling, D., B. Horton, P. Jones, T. Peterson, T. Karl, D. Parker, M. Salinger, V. Razuvayev, N.

Plummer, P. Jamason and C. Folland, 1997, “Maximum and Minimum Temperature Trends for the
Globe”, Science, 277, pp. 364–7.

Embrechts, P., A. McNeil and D. Straumann, 2000, Correlation and Dependency in Risk Management:
Properties and Pitfalls, in: M. A. H. Dempster and H. K. Moffatt (eds) Risk Management: Value at Risk
and Beyond, (Cambridge University Press).

Geman, H., 1999, “The Bermudan Triangle: Weather, Electricity and Insurance Derivatives, in:
H.Geman (ed), Insurance and Weather Derivatives, (London: Risk Books).

Hansen, J., R. Ruedy, M. Sato and R. Reynolds, 1996, “Global Surface Air Temperature in 1995:
Return to Pre-Pinatubo Level”, Geophysical Research Letters, 23, pp. 1665–8.

Huber, P. J., 1981, Robust Statistics, (New York: John Wiley & Sons).

IPCC, 2001, “Climate Change 2001-The Scientific Basis”, Technical Report, Technical Summary of the
Working Group I Report.

Jewson, S., and A. Brix, 2001, “Sunny Outlook for Weather Investors?”, Environmental Finance,
February, pp. 28-9.

Knappenberger, P., P. Michaels and R. Davis, 2001, “Nature of Observed Temperature Changes across
the United States during the 20th Century”, Climate Research, 17, pp. 45–53.

Kreps, R., 1999, “Investment Equivalent Reinsurance Pricing”, in: Proceedings of the Casualty
Actuarial Society, Volume LXXXVI, (Arlington, VA: CAS).

McCullagh, P., and J. A. Nelder, 1989, Generalized Linear Models, (London: Chapman and Hall).

Moreno, M., 2000, “Riding the Temp”, Research Paper, Speedwell Weather Derivatives, Available at:
http://www.weatherderivs.com/Papers/Riding.pdf.

CLIMATE RISK AND THE WEATHER MARKET

 150
W E AT H E R D E R I VAT I V E Nakada, P., H. Shah, H. Koyluoglu and O. Collignon, 1999, “P&C RAROC: A Catalyst for Improved
MODELLING AND Capital Management in the Property and Casualty Insurance Industry”, Journal of Risk, 1, pp. 52–70.

VA L U AT I O N : A Torró, H., V. Meneu and E. Valor, 2001, “Single Factor Stochastic Models with Seasonality Applied to
Underlying Weather Derivatives Variables”, Research Paper, Facultat d’Economia, Universitat de Valencia,
S TAT I S T I C A L
available at: http://www.efmaefm.org/htorro.pdf.
PERSPECTIVE
Wilks, D.S., and R. L. Wilby, 1999, “The Weather Generation Game: A Review of Stochastic Weather
Models”, Progress in Physical Geography, 23, pp. 329–57.

WRMA, 2002, “The Weather Risk Management Industry: Survey Findings for November 1997 to March
2001”, Report prepared for Weather Risk Management Association by PriceWaterhouseCoopers,
Available at http://www.wrma.org.

RISK BOOKS